Using Generative AI to capture High Fidelity Temporal Dynamics to target Vehicular Systems
Keywords: Generative model, Cybersecurity, Vehicle System, Synthetic Data
TL;DR: This paper presents a Long Short-Term Memory (LSTM) based Conditional Generative Adversarial Network (GAN) model, which trains on limited available real data and is then able to generate synthetic time series data mimicking the actual vehicle data
Abstract: Generative models have transformed the creation of text, images, and video content by enabling machines to generate high-quality, realistic outputs. These models are now widely being adopted in advanced fields like natural language processing, computer vision, and media production. Since vehicle data is limited due to proprietary concerns, utilizing generative models to mimic complex vehicle behaviors would provide powerful tools for creating synthetic data that can serve as a crucial component for enhancing the fidelity of vehicle models, better predictive maintenance, more robust control systems, autonomous driving features and resilient defense mechanism against cyber threats. This paper presents a Long Short-Term Memory (LSTM) based Conditional Generative Adversarial
Network (GAN) model, which trains on limited available real vehicle data and is then able to generate synthetic time series data mimicking the actual vehicle data. The LSTM network helps in learning temporal characteristics of vehicle network traffic without needing the system details, which makes it applicable to wide range of vehicle networks. The conditional layer adds auxiliary information by labeling
data for different driving scenarios for training and generating data. The quality of the synthetic data is evaluated visually and quantitatively using metrics such as Maximum Mean Discrepancy (MMD), Predictive and Discriminative Scores. For demonstration purposes, the generative model is integrated into a validated vehicle model, where it successfully generates synthetic sensor feedback corresponding to the dynamic driving scenarios. This showcases the model’s ability to simulate realistic sensor data in response to varying vehicle operations. Leveraging the high similarity to actual data, the generative model is further demonstrated for its potential use as malicious attack mechanism due to its deception capabilities against state of the art Intrusion Detection System (IDS). Without triggering the thresholds of the IDS, the model is able to penetrate the network stealthily with a low detection rate of 47.05%, compared to the 90% or higher detection rates of other known attacks. This effort is intended to serve as a test benchmark to develop more robust ML/AI based defense mechanisms.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12791
Loading